Abstract: Considering both the characteristics of clustering volatility and fat-tail of the data distribution of returns on financial assets especially the impact of conditional heteroscedaticity on the estimate of dynamic VAR, a GARCH-EVT model is developed by EVT (extreme value theory) to calculate the dynamic VAR(value at risk) of SSCI (Shanghai stock comprehensive index), then the model is compared with the GARCH-NORMAL model. The empirical analysis and posterior test results reveal that the GARCH-EVT model is superior to the GARCH-NORMAL model, because the former can solve better the problems of clustering volatility and fat-tail phenomenon. So it provides the managers and investors with quantitatively useful means for risk control.